Linux Audio

Check our new training course

Loading...
v4.10.11
   1/*
   2 * ARC Cache Management
   3 *
   4 * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
   5 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
   6 *
   7 * This program is free software; you can redistribute it and/or modify
   8 * it under the terms of the GNU General Public License version 2 as
   9 * published by the Free Software Foundation.
  10 */
  11
  12#include <linux/module.h>
  13#include <linux/mm.h>
  14#include <linux/sched.h>
  15#include <linux/cache.h>
  16#include <linux/mmu_context.h>
  17#include <linux/syscalls.h>
  18#include <linux/uaccess.h>
  19#include <linux/pagemap.h>
  20#include <asm/cacheflush.h>
  21#include <asm/cachectl.h>
  22#include <asm/setup.h>
  23
  24static int l2_line_sz;
  25static int ioc_exists;
  26int slc_enable = 1, ioc_enable = 1;
  27unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
  28unsigned long perip_end = 0xFFFFFFFF; /* legacy value */
  29
  30void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
  31			       unsigned long sz, const int cacheop);
  32
  33void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
  34void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
  35void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
  36
  37char *arc_cache_mumbojumbo(int c, char *buf, int len)
  38{
  39	int n = 0;
  40	struct cpuinfo_arc_cache *p;
  41
  42#define PR_CACHE(p, cfg, str)						\
  43	if (!(p)->line_len)						\
  44		n += scnprintf(buf + n, len - n, str"\t\t: N/A\n");	\
  45	else								\
  46		n += scnprintf(buf + n, len - n,			\
  47			str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n",	\
  48			(p)->sz_k, (p)->assoc, (p)->line_len,		\
  49			(p)->vipt ? "VIPT" : "PIPT",			\
  50			(p)->alias ? " aliasing" : "",			\
  51			IS_USED_CFG(cfg));
  52
  53	PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
  54	PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
  55
 
 
 
  56	p = &cpuinfo_arc700[c].slc;
  57	if (p->line_len)
  58		n += scnprintf(buf + n, len - n,
  59			       "SLC\t\t: %uK, %uB Line%s\n",
  60			       p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
  61
  62	n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
  63		       perip_base,
  64		       IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency "));
  65
  66	return buf;
  67}
  68
  69/*
  70 * Read the Cache Build Confuration Registers, Decode them and save into
  71 * the cpuinfo structure for later use.
  72 * No Validation done here, simply read/convert the BCRs
  73 */
  74static void read_decode_cache_bcr_arcv2(int cpu)
  75{
  76	struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
 
  77	struct bcr_generic sbcr;
  78
  79	struct bcr_slc_cfg {
  80#ifdef CONFIG_CPU_BIG_ENDIAN
  81		unsigned int pad:24, way:2, lsz:2, sz:4;
  82#else
  83		unsigned int sz:4, lsz:2, way:2, pad:24;
  84#endif
  85	} slc_cfg;
  86
  87	struct bcr_clust_cfg {
  88#ifdef CONFIG_CPU_BIG_ENDIAN
  89		unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
  90#else
  91		unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
  92#endif
  93	} cbcr;
  94
  95	struct bcr_volatile {
  96#ifdef CONFIG_CPU_BIG_ENDIAN
  97		unsigned int start:4, limit:4, pad:22, order:1, disable:1;
  98#else
  99		unsigned int disable:1, order:1, pad:22, limit:4, start:4;
 100#endif
 101	} vol;
 102
 103
 104	READ_BCR(ARC_REG_SLC_BCR, sbcr);
 105	if (sbcr.ver) {
 106		READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
 
 107		p_slc->sz_k = 128 << slc_cfg.sz;
 108		l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
 109	}
 110
 111	READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
 112	if (cbcr.c)
 113		ioc_exists = 1;
 114	else
 115		ioc_enable = 0;
 116
 117	/* HS 2.0 didn't have AUX_VOL */
 118	if (cpuinfo_arc700[cpu].core.family > 0x51) {
 119		READ_BCR(AUX_VOL, vol);
 120		perip_base = vol.start << 28;
 121		/* HS 3.0 has limit and strict-ordering fields */
 122		if (cpuinfo_arc700[cpu].core.family > 0x52)
 123			perip_end = (vol.limit << 28) - 1;
 124	}
 125}
 126
 127void read_decode_cache_bcr(void)
 128{
 129	struct cpuinfo_arc_cache *p_ic, *p_dc;
 130	unsigned int cpu = smp_processor_id();
 131	struct bcr_cache {
 132#ifdef CONFIG_CPU_BIG_ENDIAN
 133		unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
 134#else
 135		unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
 136#endif
 137	} ibcr, dbcr;
 138
 139	p_ic = &cpuinfo_arc700[cpu].icache;
 140	READ_BCR(ARC_REG_IC_BCR, ibcr);
 141
 142	if (!ibcr.ver)
 143		goto dc_chk;
 144
 145	if (ibcr.ver <= 3) {
 146		BUG_ON(ibcr.config != 3);
 147		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
 148	} else if (ibcr.ver >= 4) {
 149		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
 150	}
 151
 152	p_ic->line_len = 8 << ibcr.line_len;
 153	p_ic->sz_k = 1 << (ibcr.sz - 1);
 
 154	p_ic->vipt = 1;
 155	p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
 156
 157dc_chk:
 158	p_dc = &cpuinfo_arc700[cpu].dcache;
 159	READ_BCR(ARC_REG_DC_BCR, dbcr);
 160
 161	if (!dbcr.ver)
 162		goto slc_chk;
 163
 164	if (dbcr.ver <= 3) {
 165		BUG_ON(dbcr.config != 2);
 166		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
 167		p_dc->vipt = 1;
 168		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
 169	} else if (dbcr.ver >= 4) {
 170		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
 171		p_dc->vipt = 0;
 172		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
 173	}
 174
 175	p_dc->line_len = 16 << dbcr.line_len;
 176	p_dc->sz_k = 1 << (dbcr.sz - 1);
 
 177
 178slc_chk:
 179	if (is_isa_arcv2())
 180                read_decode_cache_bcr_arcv2(cpu);
 181}
 182
 183/*
 184 * Line Operation on {I,D}-Cache
 185 */
 186
 187#define OP_INV		0x1
 188#define OP_FLUSH	0x2
 189#define OP_FLUSH_N_INV	0x3
 190#define OP_INV_IC	0x4
 191
 192/*
 193 *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
 194 *
 195 * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
 196 * The orig Cache Management Module "CDU" only required paddr to invalidate a
 197 * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
 198 * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
 199 * the exact same line.
 200 *
 201 * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
 202 * paddr alone could not be used to correctly index the cache.
 203 *
 204 * ------------------
 205 * MMU v1/v2 (Fixed Page Size 8k)
 206 * ------------------
 207 * The solution was to provide CDU with these additonal vaddr bits. These
 208 * would be bits [x:13], x would depend on cache-geometry, 13 comes from
 209 * standard page size of 8k.
 210 * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
 211 * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
 212 * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
 213 * represent the offset within cache-line. The adv of using this "clumsy"
 214 * interface for additional info was no new reg was needed in CDU programming
 215 * model.
 216 *
 217 * 17:13 represented the max num of bits passable, actual bits needed were
 218 * fewer, based on the num-of-aliases possible.
 219 * -for 2 alias possibility, only bit 13 needed (32K cache)
 220 * -for 4 alias possibility, bits 14:13 needed (64K cache)
 221 *
 222 * ------------------
 223 * MMU v3
 224 * ------------------
 225 * This ver of MMU supports variable page sizes (1k-16k): although Linux will
 226 * only support 8k (default), 16k and 4k.
 227 * However from hardware perspective, smaller page sizes aggravate aliasing
 228 * meaning more vaddr bits needed to disambiguate the cache-line-op ;
 229 * the existing scheme of piggybacking won't work for certain configurations.
 230 * Two new registers IC_PTAG and DC_PTAG inttoduced.
 231 * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
 232 */
 233
 234static inline
 235void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
 236			  unsigned long sz, const int op)
 237{
 238	unsigned int aux_cmd;
 239	int num_lines;
 240	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 241
 242	if (op == OP_INV_IC) {
 243		aux_cmd = ARC_REG_IC_IVIL;
 244	} else {
 245		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
 246		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 247	}
 248
 249	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 250	 * and have @paddr - aligned to cache line and integral @num_lines.
 251	 * This however can be avoided for page sized since:
 252	 *  -@paddr will be cache-line aligned already (being page aligned)
 253	 *  -@sz will be integral multiple of line size (being page sized).
 254	 */
 255	if (!full_page) {
 256		sz += paddr & ~CACHE_LINE_MASK;
 257		paddr &= CACHE_LINE_MASK;
 258		vaddr &= CACHE_LINE_MASK;
 259	}
 260
 261	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 262
 263	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
 264	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
 265
 266	while (num_lines-- > 0) {
 267		write_aux_reg(aux_cmd, paddr);
 268		paddr += L1_CACHE_BYTES;
 269	}
 270}
 271
 272/*
 273 * For ARC700 MMUv3 I-cache and D-cache flushes
 274 *  - ARC700 programming model requires paddr and vaddr be passed in seperate
 275 *    AUX registers (*_IV*L and *_PTAG respectively) irrespective of whether the
 276 *    caches actually alias or not.
 277 * -  For HS38, only the aliasing I-cache configuration uses the PTAG reg
 278 *    (non aliasing I-cache version doesn't; while D-cache can't possibly alias)
 279 */
 280static inline
 281void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
 282			  unsigned long sz, const int op)
 283{
 284	unsigned int aux_cmd, aux_tag;
 285	int num_lines;
 286	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 287
 288	if (op == OP_INV_IC) {
 289		aux_cmd = ARC_REG_IC_IVIL;
 290		aux_tag = ARC_REG_IC_PTAG;
 291	} else {
 292		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 293		aux_tag = ARC_REG_DC_PTAG;
 294	}
 295
 296	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 297	 * and have @paddr - aligned to cache line and integral @num_lines.
 298	 * This however can be avoided for page sized since:
 299	 *  -@paddr will be cache-line aligned already (being page aligned)
 300	 *  -@sz will be integral multiple of line size (being page sized).
 301	 */
 302	if (!full_page) {
 303		sz += paddr & ~CACHE_LINE_MASK;
 304		paddr &= CACHE_LINE_MASK;
 305		vaddr &= CACHE_LINE_MASK;
 306	}
 307	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 308
 309	/*
 310	 * MMUv3, cache ops require paddr in PTAG reg
 311	 * if V-P const for loop, PTAG can be written once outside loop
 312	 */
 313	if (full_page)
 314		write_aux_reg(aux_tag, paddr);
 315
 316	/*
 317	 * This is technically for MMU v4, using the MMU v3 programming model
 318	 * Special work for HS38 aliasing I-cache configuration with PAE40
 319	 *   - upper 8 bits of paddr need to be written into PTAG_HI
 320	 *   - (and needs to be written before the lower 32 bits)
 321	 * Note that PTAG_HI is hoisted outside the line loop
 322	 */
 323	if (is_pae40_enabled() && op == OP_INV_IC)
 324		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
 325
 326	while (num_lines-- > 0) {
 327		if (!full_page) {
 328			write_aux_reg(aux_tag, paddr);
 329			paddr += L1_CACHE_BYTES;
 330		}
 331
 332		write_aux_reg(aux_cmd, vaddr);
 333		vaddr += L1_CACHE_BYTES;
 334	}
 335}
 336
 337/*
 338 * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
 339 * Here's how cache ops are implemented
 340 *
 341 *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
 342 *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
 343 *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
 344 *    respectively, similar to MMU v3 programming model, hence
 345 *    __cache_line_loop_v3() is used)
 346 *
 347 * If PAE40 is enabled, independent of aliasing considerations, the higher bits
 348 * needs to be written into PTAG_HI
 349 */
 350static inline
 351void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
 352			  unsigned long sz, const int cacheop)
 353{
 354	unsigned int aux_cmd;
 355	int num_lines;
 356	const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 357
 358	if (cacheop == OP_INV_IC) {
 359		aux_cmd = ARC_REG_IC_IVIL;
 360	} else {
 361		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
 362		aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 363	}
 364
 365	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 366	 * and have @paddr - aligned to cache line and integral @num_lines.
 367	 * This however can be avoided for page sized since:
 368	 *  -@paddr will be cache-line aligned already (being page aligned)
 369	 *  -@sz will be integral multiple of line size (being page sized).
 370	 */
 371	if (!full_page_op) {
 372		sz += paddr & ~CACHE_LINE_MASK;
 373		paddr &= CACHE_LINE_MASK;
 374	}
 375
 376	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 377
 378	/*
 379	 * For HS38 PAE40 configuration
 380	 *   - upper 8 bits of paddr need to be written into PTAG_HI
 381	 *   - (and needs to be written before the lower 32 bits)
 382	 */
 383	if (is_pae40_enabled()) {
 384		if (cacheop == OP_INV_IC)
 385			/*
 386			 * Non aliasing I-cache in HS38,
 387			 * aliasing I-cache handled in __cache_line_loop_v3()
 388			 */
 389			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
 390		else
 391			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
 392	}
 393
 394	while (num_lines-- > 0) {
 395		write_aux_reg(aux_cmd, paddr);
 396		paddr += L1_CACHE_BYTES;
 397	}
 398}
 399
 400#if (CONFIG_ARC_MMU_VER < 3)
 401#define __cache_line_loop	__cache_line_loop_v2
 402#elif (CONFIG_ARC_MMU_VER == 3)
 403#define __cache_line_loop	__cache_line_loop_v3
 404#elif (CONFIG_ARC_MMU_VER > 3)
 405#define __cache_line_loop	__cache_line_loop_v4
 406#endif
 407
 408#ifdef CONFIG_ARC_HAS_DCACHE
 409
 410/***************************************************************
 411 * Machine specific helpers for Entire D-Cache or Per Line ops
 412 */
 413
 414static inline void __before_dc_op(const int op)
 415{
 416	if (op == OP_FLUSH_N_INV) {
 417		/* Dcache provides 2 cmd: FLUSH or INV
 418		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
 419		 * flush-n-inv is achieved by INV cmd but with IM=1
 420		 * So toggle INV sub-mode depending on op request and default
 421		 */
 422		const unsigned int ctl = ARC_REG_DC_CTRL;
 423		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
 424	}
 425}
 426
 427static inline void __after_dc_op(const int op)
 428{
 429	if (op & OP_FLUSH) {
 430		const unsigned int ctl = ARC_REG_DC_CTRL;
 431		unsigned int reg;
 432
 433		/* flush / flush-n-inv both wait */
 434		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
 435			;
 436
 437		/* Switch back to default Invalidate mode */
 438		if (op == OP_FLUSH_N_INV)
 439			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
 440	}
 441}
 442
 443/*
 444 * Operation on Entire D-Cache
 445 * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
 446 * Note that constant propagation ensures all the checks are gone
 447 * in generated code
 448 */
 449static inline void __dc_entire_op(const int op)
 450{
 451	int aux;
 452
 453	__before_dc_op(op);
 454
 455	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
 456		aux = ARC_REG_DC_IVDC;
 457	else
 458		aux = ARC_REG_DC_FLSH;
 459
 460	write_aux_reg(aux, 0x1);
 461
 462	__after_dc_op(op);
 463}
 464
 465static inline void __dc_disable(void)
 466{
 467	const int r = ARC_REG_DC_CTRL;
 468
 469	__dc_entire_op(OP_FLUSH_N_INV);
 470	write_aux_reg(r, read_aux_reg(r) | DC_CTRL_DIS);
 471}
 472
 473static void __dc_enable(void)
 474{
 475	const int r = ARC_REG_DC_CTRL;
 476
 477	write_aux_reg(r, read_aux_reg(r) & ~DC_CTRL_DIS);
 478}
 479
 480/* For kernel mappings cache operation: index is same as paddr */
 481#define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
 482
 483/*
 484 * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
 485 */
 486static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
 487				unsigned long sz, const int op)
 488{
 489	unsigned long flags;
 490
 491	local_irq_save(flags);
 492
 493	__before_dc_op(op);
 494
 495	__cache_line_loop(paddr, vaddr, sz, op);
 496
 497	__after_dc_op(op);
 498
 499	local_irq_restore(flags);
 500}
 501
 502#else
 503
 504#define __dc_entire_op(op)
 505#define __dc_disable()
 506#define __dc_enable()
 507#define __dc_line_op(paddr, vaddr, sz, op)
 508#define __dc_line_op_k(paddr, sz, op)
 509
 510#endif /* CONFIG_ARC_HAS_DCACHE */
 511
 512#ifdef CONFIG_ARC_HAS_ICACHE
 513
 514static inline void __ic_entire_inv(void)
 515{
 516	write_aux_reg(ARC_REG_IC_IVIC, 1);
 517	read_aux_reg(ARC_REG_IC_CTRL);	/* blocks */
 518}
 519
 520static inline void
 521__ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
 522			  unsigned long sz)
 523{
 524	unsigned long flags;
 525
 526	local_irq_save(flags);
 527	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC);
 528	local_irq_restore(flags);
 529}
 530
 531#ifndef CONFIG_SMP
 532
 533#define __ic_line_inv_vaddr(p, v, s)	__ic_line_inv_vaddr_local(p, v, s)
 534
 535#else
 536
 537struct ic_inv_args {
 538	phys_addr_t paddr, vaddr;
 539	int sz;
 540};
 541
 542static void __ic_line_inv_vaddr_helper(void *info)
 543{
 544        struct ic_inv_args *ic_inv = info;
 545
 546        __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
 547}
 548
 549static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
 550				unsigned long sz)
 551{
 552	struct ic_inv_args ic_inv = {
 553		.paddr = paddr,
 554		.vaddr = vaddr,
 555		.sz    = sz
 556	};
 557
 558	on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
 559}
 560
 561#endif	/* CONFIG_SMP */
 562
 563#else	/* !CONFIG_ARC_HAS_ICACHE */
 564
 565#define __ic_entire_inv()
 566#define __ic_line_inv_vaddr(pstart, vstart, sz)
 567
 568#endif /* CONFIG_ARC_HAS_ICACHE */
 569
 570noinline void slc_op(phys_addr_t paddr, unsigned long sz, const int op)
 571{
 572#ifdef CONFIG_ISA_ARCV2
 573	/*
 574	 * SLC is shared between all cores and concurrent aux operations from
 575	 * multiple cores need to be serialized using a spinlock
 576	 * A concurrent operation can be silently ignored and/or the old/new
 577	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
 578	 * below)
 579	 */
 580	static DEFINE_SPINLOCK(lock);
 581	unsigned long flags;
 582	unsigned int ctrl;
 583
 584	spin_lock_irqsave(&lock, flags);
 585
 586	/*
 587	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
 588	 *  - b'000 (default) is Flush,
 589	 *  - b'001 is Invalidate if CTRL.IM == 0
 590	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
 591	 */
 592	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
 593
 594	/* Don't rely on default value of IM bit */
 595	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
 596		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
 597	else
 598		ctrl |= SLC_CTRL_IM;
 599
 600	if (op & OP_INV)
 601		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
 602	else
 603		ctrl &= ~SLC_CTRL_RGN_OP_INV;
 604
 605	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
 606
 607	/*
 608	 * Lower bits are ignored, no need to clip
 609	 * END needs to be setup before START (latter triggers the operation)
 610	 * END can't be same as START, so add (l2_line_sz - 1) to sz
 611	 */
 612	write_aux_reg(ARC_REG_SLC_RGN_END, (paddr + sz + l2_line_sz - 1));
 613	write_aux_reg(ARC_REG_SLC_RGN_START, paddr);
 614
 615	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
 616
 617	spin_unlock_irqrestore(&lock, flags);
 618#endif
 619}
 620
 621noinline static void slc_entire_op(const int op)
 622{
 623	unsigned int ctrl, r = ARC_REG_SLC_CTRL;
 624
 625	ctrl = read_aux_reg(r);
 626
 627	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
 628		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
 629	else
 630		ctrl |= SLC_CTRL_IM;
 631
 632	write_aux_reg(r, ctrl);
 633
 634	write_aux_reg(ARC_REG_SLC_INVALIDATE, 1);
 635
 636	/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
 637	read_aux_reg(r);
 638
 639	/* Important to wait for flush to complete */
 640	while (read_aux_reg(r) & SLC_CTRL_BUSY);
 641}
 642
 643static inline void arc_slc_disable(void)
 644{
 645	const int r = ARC_REG_SLC_CTRL;
 646
 647	slc_entire_op(OP_FLUSH_N_INV);
 648	write_aux_reg(r, read_aux_reg(r) | SLC_CTRL_DIS);
 649}
 650
 651static inline void arc_slc_enable(void)
 652{
 653	const int r = ARC_REG_SLC_CTRL;
 654
 655	write_aux_reg(r, read_aux_reg(r) & ~SLC_CTRL_DIS);
 656}
 657
 658/***********************************************************
 659 * Exported APIs
 660 */
 661
 662/*
 663 * Handle cache congruency of kernel and userspace mappings of page when kernel
 664 * writes-to/reads-from
 665 *
 666 * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
 667 *  -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
 668 *  -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
 669 *  -In SMP, if hardware caches are coherent
 670 *
 671 * There's a corollary case, where kernel READs from a userspace mapped page.
 672 * If the U-mapping is not congruent to to K-mapping, former needs flushing.
 673 */
 674void flush_dcache_page(struct page *page)
 675{
 676	struct address_space *mapping;
 677
 678	if (!cache_is_vipt_aliasing()) {
 679		clear_bit(PG_dc_clean, &page->flags);
 680		return;
 681	}
 682
 683	/* don't handle anon pages here */
 684	mapping = page_mapping(page);
 685	if (!mapping)
 686		return;
 687
 688	/*
 689	 * pagecache page, file not yet mapped to userspace
 690	 * Make a note that K-mapping is dirty
 691	 */
 692	if (!mapping_mapped(mapping)) {
 693		clear_bit(PG_dc_clean, &page->flags);
 694	} else if (page_mapcount(page)) {
 695
 696		/* kernel reading from page with U-mapping */
 697		phys_addr_t paddr = (unsigned long)page_address(page);
 698		unsigned long vaddr = page->index << PAGE_SHIFT;
 699
 700		if (addr_not_cache_congruent(paddr, vaddr))
 701			__flush_dcache_page(paddr, vaddr);
 702	}
 703}
 704EXPORT_SYMBOL(flush_dcache_page);
 705
 706/*
 707 * DMA ops for systems with L1 cache only
 708 * Make memory coherent with L1 cache by flushing/invalidating L1 lines
 709 */
 710static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
 711{
 712	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
 713}
 714
 715static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
 716{
 717	__dc_line_op_k(start, sz, OP_INV);
 718}
 719
 720static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
 721{
 722	__dc_line_op_k(start, sz, OP_FLUSH);
 723}
 724
 725/*
 726 * DMA ops for systems with both L1 and L2 caches, but without IOC
 727 * Both L1 and L2 lines need to be explicitly flushed/invalidated
 728 */
 729static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
 730{
 731	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
 732	slc_op(start, sz, OP_FLUSH_N_INV);
 733}
 734
 735static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
 736{
 737	__dc_line_op_k(start, sz, OP_INV);
 738	slc_op(start, sz, OP_INV);
 739}
 740
 741static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
 742{
 743	__dc_line_op_k(start, sz, OP_FLUSH);
 744	slc_op(start, sz, OP_FLUSH);
 745}
 746
 747/*
 748 * DMA ops for systems with IOC
 749 * IOC hardware snoops all DMA traffic keeping the caches consistent with
 750 * memory - eliding need for any explicit cache maintenance of DMA buffers
 751 */
 752static void __dma_cache_wback_inv_ioc(phys_addr_t start, unsigned long sz) {}
 753static void __dma_cache_inv_ioc(phys_addr_t start, unsigned long sz) {}
 754static void __dma_cache_wback_ioc(phys_addr_t start, unsigned long sz) {}
 755
 756/*
 757 * Exported DMA API
 758 */
 759void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
 760{
 761	__dma_cache_wback_inv(start, sz);
 762}
 763EXPORT_SYMBOL(dma_cache_wback_inv);
 764
 765void dma_cache_inv(phys_addr_t start, unsigned long sz)
 766{
 767	__dma_cache_inv(start, sz);
 768}
 769EXPORT_SYMBOL(dma_cache_inv);
 770
 771void dma_cache_wback(phys_addr_t start, unsigned long sz)
 772{
 773	__dma_cache_wback(start, sz);
 774}
 775EXPORT_SYMBOL(dma_cache_wback);
 776
 777/*
 778 * This is API for making I/D Caches consistent when modifying
 779 * kernel code (loadable modules, kprobes, kgdb...)
 780 * This is called on insmod, with kernel virtual address for CODE of
 781 * the module. ARC cache maintenance ops require PHY address thus we
 782 * need to convert vmalloc addr to PHY addr
 783 */
 784void flush_icache_range(unsigned long kstart, unsigned long kend)
 785{
 786	unsigned int tot_sz;
 787
 788	WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
 789
 790	/* Shortcut for bigger flush ranges.
 791	 * Here we don't care if this was kernel virtual or phy addr
 792	 */
 793	tot_sz = kend - kstart;
 794	if (tot_sz > PAGE_SIZE) {
 795		flush_cache_all();
 796		return;
 797	}
 798
 799	/* Case: Kernel Phy addr (0x8000_0000 onwards) */
 800	if (likely(kstart > PAGE_OFFSET)) {
 801		/*
 802		 * The 2nd arg despite being paddr will be used to index icache
 803		 * This is OK since no alternate virtual mappings will exist
 804		 * given the callers for this case: kprobe/kgdb in built-in
 805		 * kernel code only.
 806		 */
 807		__sync_icache_dcache(kstart, kstart, kend - kstart);
 808		return;
 809	}
 810
 811	/*
 812	 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
 813	 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
 814	 *     handling of kernel vaddr.
 815	 *
 816	 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
 817	 *     it still needs to handle  a 2 page scenario, where the range
 818	 *     straddles across 2 virtual pages and hence need for loop
 819	 */
 820	while (tot_sz > 0) {
 821		unsigned int off, sz;
 822		unsigned long phy, pfn;
 823
 824		off = kstart % PAGE_SIZE;
 825		pfn = vmalloc_to_pfn((void *)kstart);
 826		phy = (pfn << PAGE_SHIFT) + off;
 827		sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
 828		__sync_icache_dcache(phy, kstart, sz);
 829		kstart += sz;
 830		tot_sz -= sz;
 831	}
 832}
 833EXPORT_SYMBOL(flush_icache_range);
 834
 835/*
 836 * General purpose helper to make I and D cache lines consistent.
 837 * @paddr is phy addr of region
 838 * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
 839 *    However in one instance, when called by kprobe (for a breakpt in
 840 *    builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
 841 *    use a paddr to index the cache (despite VIPT). This is fine since since a
 842 *    builtin kernel page will not have any virtual mappings.
 843 *    kprobe on loadable module will be kernel vaddr.
 844 */
 845void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
 846{
 847	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
 848	__ic_line_inv_vaddr(paddr, vaddr, len);
 849}
 850
 851/* wrapper to compile time eliminate alignment checks in flush loop */
 852void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
 853{
 854	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
 855}
 856
 857/*
 858 * wrapper to clearout kernel or userspace mappings of a page
 859 * For kernel mappings @vaddr == @paddr
 860 */
 861void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
 862{
 863	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
 864}
 865
 866noinline void flush_cache_all(void)
 867{
 868	unsigned long flags;
 869
 870	local_irq_save(flags);
 871
 872	__ic_entire_inv();
 873	__dc_entire_op(OP_FLUSH_N_INV);
 874
 875	local_irq_restore(flags);
 876
 877}
 878
 879#ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
 880
 881void flush_cache_mm(struct mm_struct *mm)
 882{
 883	flush_cache_all();
 884}
 885
 886void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
 887		      unsigned long pfn)
 888{
 889	unsigned int paddr = pfn << PAGE_SHIFT;
 890
 891	u_vaddr &= PAGE_MASK;
 892
 893	__flush_dcache_page(paddr, u_vaddr);
 894
 895	if (vma->vm_flags & VM_EXEC)
 896		__inv_icache_page(paddr, u_vaddr);
 897}
 898
 899void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 900		       unsigned long end)
 901{
 902	flush_cache_all();
 903}
 904
 905void flush_anon_page(struct vm_area_struct *vma, struct page *page,
 906		     unsigned long u_vaddr)
 907{
 908	/* TBD: do we really need to clear the kernel mapping */
 909	__flush_dcache_page(page_address(page), u_vaddr);
 910	__flush_dcache_page(page_address(page), page_address(page));
 911
 912}
 913
 914#endif
 915
 916void copy_user_highpage(struct page *to, struct page *from,
 917	unsigned long u_vaddr, struct vm_area_struct *vma)
 918{
 919	void *kfrom = kmap_atomic(from);
 920	void *kto = kmap_atomic(to);
 921	int clean_src_k_mappings = 0;
 922
 923	/*
 924	 * If SRC page was already mapped in userspace AND it's U-mapping is
 925	 * not congruent with K-mapping, sync former to physical page so that
 926	 * K-mapping in memcpy below, sees the right data
 927	 *
 928	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
 929	 * equally valid for SRC page as well
 930	 *
 931	 * For !VIPT cache, all of this gets compiled out as
 932	 * addr_not_cache_congruent() is 0
 933	 */
 934	if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
 935		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
 936		clean_src_k_mappings = 1;
 937	}
 938
 939	copy_page(kto, kfrom);
 940
 941	/*
 942	 * Mark DST page K-mapping as dirty for a later finalization by
 943	 * update_mmu_cache(). Although the finalization could have been done
 944	 * here as well (given that both vaddr/paddr are available).
 945	 * But update_mmu_cache() already has code to do that for other
 946	 * non copied user pages (e.g. read faults which wire in pagecache page
 947	 * directly).
 948	 */
 949	clear_bit(PG_dc_clean, &to->flags);
 950
 951	/*
 952	 * if SRC was already usermapped and non-congruent to kernel mapping
 953	 * sync the kernel mapping back to physical page
 954	 */
 955	if (clean_src_k_mappings) {
 956		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
 957		set_bit(PG_dc_clean, &from->flags);
 958	} else {
 959		clear_bit(PG_dc_clean, &from->flags);
 960	}
 961
 962	kunmap_atomic(kto);
 963	kunmap_atomic(kfrom);
 964}
 965
 966void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
 967{
 968	clear_page(to);
 969	clear_bit(PG_dc_clean, &page->flags);
 970}
 971
 972
 973/**********************************************************************
 974 * Explicit Cache flush request from user space via syscall
 975 * Needed for JITs which generate code on the fly
 976 */
 977SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
 978{
 979	/* TBD: optimize this */
 980	flush_cache_all();
 981	return 0;
 982}
 983
 984/*
 985 * IO-Coherency (IOC) setup rules:
 986 *
 987 * 1. Needs to be at system level, so only once by Master core
 988 *    Non-Masters need not be accessing caches at that time
 989 *    - They are either HALT_ON_RESET and kick started much later or
 990 *    - if run on reset, need to ensure that arc_platform_smp_wait_to_boot()
 991 *      doesn't perturb caches or coherency unit
 992 *
 993 * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC,
 994 *    otherwise any straggler data might behave strangely post IOC enabling
 995 *
 996 * 3. All Caches need to be disabled when setting up IOC to elide any in-flight
 997 *    Coherency transactions
 998 */
 999noinline void __init arc_ioc_setup(void)
1000{
1001	unsigned int ap_sz;
1002
1003	/* Flush + invalidate + disable L1 dcache */
1004	__dc_disable();
1005
1006	/* Flush + invalidate SLC */
1007	if (read_aux_reg(ARC_REG_SLC_BCR))
1008		slc_entire_op(OP_FLUSH_N_INV);
1009
1010	/* IOC Aperture start: TDB: handle non default CONFIG_LINUX_LINK_BASE */
1011	write_aux_reg(ARC_REG_IO_COH_AP0_BASE, 0x80000);
1012
1013	/*
1014	 * IOC Aperture size:
1015	 *   decoded as 2 ^ (SIZE + 2) KB: so setting 0x11 implies 512M
1016	 * TBD: fix for PGU + 1GB of low mem
1017	 * TBD: fix for PAE
1018	 */
1019	ap_sz = order_base_2(arc_get_mem_sz()/1024) - 2;
1020	write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, ap_sz);
1021
1022	write_aux_reg(ARC_REG_IO_COH_PARTIAL, 1);
1023	write_aux_reg(ARC_REG_IO_COH_ENABLE, 1);
1024
1025	/* Re-enable L1 dcache */
1026	__dc_enable();
1027}
1028
1029void __init arc_cache_init_master(void)
1030{
1031	unsigned int __maybe_unused cpu = smp_processor_id();
 
 
 
1032
1033	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
1034		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
1035
1036		if (!ic->line_len)
1037			panic("cache support enabled but non-existent cache\n");
1038
1039		if (ic->line_len != L1_CACHE_BYTES)
1040			panic("ICache line [%d] != kernel Config [%d]",
1041			      ic->line_len, L1_CACHE_BYTES);
1042
 
 
 
 
1043		/*
1044		 * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG
1045		 * pair to provide vaddr/paddr respectively, just as in MMU v3
1046		 */
1047		if (is_isa_arcv2() && ic->alias)
1048			_cache_line_loop_ic_fn = __cache_line_loop_v3;
1049		else
1050			_cache_line_loop_ic_fn = __cache_line_loop;
1051	}
1052
1053	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
1054		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
1055
1056		if (!dc->line_len)
1057			panic("cache support enabled but non-existent cache\n");
1058
1059		if (dc->line_len != L1_CACHE_BYTES)
1060			panic("DCache line [%d] != kernel Config [%d]",
1061			      dc->line_len, L1_CACHE_BYTES);
1062
1063		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
1064		if (is_isa_arcompact()) {
1065			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
1066			int num_colors = dc->sz_k/dc->assoc/TO_KB(PAGE_SIZE);
1067
1068			if (dc->alias) {
1069				if (!handled)
1070					panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1071				if (CACHE_COLORS_NUM != num_colors)
1072					panic("CACHE_COLORS_NUM not optimized for config\n");
1073			} else if (!dc->alias && handled) {
1074				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
1075			}
1076		}
1077	}
1078
1079	/* Note that SLC disable not formally supported till HS 3.0 */
1080	if (is_isa_arcv2() && l2_line_sz && !slc_enable)
1081		arc_slc_disable();
1082
1083	if (is_isa_arcv2() && ioc_enable)
1084		arc_ioc_setup();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1085
1086	if (is_isa_arcv2() && ioc_enable) {
1087		__dma_cache_wback_inv = __dma_cache_wback_inv_ioc;
1088		__dma_cache_inv = __dma_cache_inv_ioc;
1089		__dma_cache_wback = __dma_cache_wback_ioc;
1090	} else if (is_isa_arcv2() && l2_line_sz && slc_enable) {
1091		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
1092		__dma_cache_inv = __dma_cache_inv_slc;
1093		__dma_cache_wback = __dma_cache_wback_slc;
1094	} else {
1095		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1096		__dma_cache_inv = __dma_cache_inv_l1;
1097		__dma_cache_wback = __dma_cache_wback_l1;
1098	}
1099}
1100
1101void __ref arc_cache_init(void)
1102{
1103	unsigned int __maybe_unused cpu = smp_processor_id();
1104	char str[256];
1105
1106	printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
1107
1108	/*
1109	 * Only master CPU needs to execute rest of function:
1110	 *  - Assume SMP so all cores will have same cache config so
1111	 *    any geomtry checks will be same for all
1112	 *  - IOC setup / dma callbacks only need to be setup once
1113	 */
1114	if (!cpu)
1115		arc_cache_init_master();
1116}
v4.6
   1/*
   2 * ARC Cache Management
   3 *
   4 * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
   5 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
   6 *
   7 * This program is free software; you can redistribute it and/or modify
   8 * it under the terms of the GNU General Public License version 2 as
   9 * published by the Free Software Foundation.
  10 */
  11
  12#include <linux/module.h>
  13#include <linux/mm.h>
  14#include <linux/sched.h>
  15#include <linux/cache.h>
  16#include <linux/mmu_context.h>
  17#include <linux/syscalls.h>
  18#include <linux/uaccess.h>
  19#include <linux/pagemap.h>
  20#include <asm/cacheflush.h>
  21#include <asm/cachectl.h>
  22#include <asm/setup.h>
  23
  24static int l2_line_sz;
  25int ioc_exists;
  26volatile int slc_enable = 1, ioc_enable = 1;
  27unsigned long perip_base = ARC_UNCACHED_ADDR_SPACE; /* legacy value for boot */
 
  28
  29void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
  30			       unsigned long sz, const int cacheop);
  31
  32void (*__dma_cache_wback_inv)(phys_addr_t start, unsigned long sz);
  33void (*__dma_cache_inv)(phys_addr_t start, unsigned long sz);
  34void (*__dma_cache_wback)(phys_addr_t start, unsigned long sz);
  35
  36char *arc_cache_mumbojumbo(int c, char *buf, int len)
  37{
  38	int n = 0;
  39	struct cpuinfo_arc_cache *p;
  40
  41#define PR_CACHE(p, cfg, str)						\
  42	if (!(p)->ver)							\
  43		n += scnprintf(buf + n, len - n, str"\t\t: N/A\n");	\
  44	else								\
  45		n += scnprintf(buf + n, len - n,			\
  46			str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n",	\
  47			(p)->sz_k, (p)->assoc, (p)->line_len,		\
  48			(p)->vipt ? "VIPT" : "PIPT",			\
  49			(p)->alias ? " aliasing" : "",			\
  50			IS_USED_CFG(cfg));
  51
  52	PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
  53	PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
  54
  55	if (!is_isa_arcv2())
  56                return buf;
  57
  58	p = &cpuinfo_arc700[c].slc;
  59	if (p->ver)
  60		n += scnprintf(buf + n, len - n,
  61			       "SLC\t\t: %uK, %uB Line%s\n",
  62			       p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
  63
  64	if (ioc_exists)
  65		n += scnprintf(buf + n, len - n, "IOC\t\t:%s\n",
  66				IS_DISABLED_RUN(ioc_enable));
  67
  68	return buf;
  69}
  70
  71/*
  72 * Read the Cache Build Confuration Registers, Decode them and save into
  73 * the cpuinfo structure for later use.
  74 * No Validation done here, simply read/convert the BCRs
  75 */
  76static void read_decode_cache_bcr_arcv2(int cpu)
  77{
  78	struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
  79	struct bcr_generic uncached_space;
  80	struct bcr_generic sbcr;
  81
  82	struct bcr_slc_cfg {
  83#ifdef CONFIG_CPU_BIG_ENDIAN
  84		unsigned int pad:24, way:2, lsz:2, sz:4;
  85#else
  86		unsigned int sz:4, lsz:2, way:2, pad:24;
  87#endif
  88	} slc_cfg;
  89
  90	struct bcr_clust_cfg {
  91#ifdef CONFIG_CPU_BIG_ENDIAN
  92		unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
  93#else
  94		unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
  95#endif
  96	} cbcr;
  97
 
 
 
 
 
 
 
 
 
  98	READ_BCR(ARC_REG_SLC_BCR, sbcr);
  99	if (sbcr.ver) {
 100		READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
 101		p_slc->ver = sbcr.ver;
 102		p_slc->sz_k = 128 << slc_cfg.sz;
 103		l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
 104	}
 105
 106	READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
 107	if (cbcr.c && ioc_enable)
 108		ioc_exists = 1;
 
 
 109
 110	/* Legacy Data Uncached BCR is deprecated from v3 onwards */
 111	READ_BCR(ARC_REG_D_UNCACH_BCR, uncached_space);
 112	if (uncached_space.ver > 2)
 113		perip_base = read_aux_reg(AUX_NON_VOL) & 0xF0000000;
 
 
 
 
 114}
 115
 116void read_decode_cache_bcr(void)
 117{
 118	struct cpuinfo_arc_cache *p_ic, *p_dc;
 119	unsigned int cpu = smp_processor_id();
 120	struct bcr_cache {
 121#ifdef CONFIG_CPU_BIG_ENDIAN
 122		unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
 123#else
 124		unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
 125#endif
 126	} ibcr, dbcr;
 127
 128	p_ic = &cpuinfo_arc700[cpu].icache;
 129	READ_BCR(ARC_REG_IC_BCR, ibcr);
 130
 131	if (!ibcr.ver)
 132		goto dc_chk;
 133
 134	if (ibcr.ver <= 3) {
 135		BUG_ON(ibcr.config != 3);
 136		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
 137	} else if (ibcr.ver >= 4) {
 138		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
 139	}
 140
 141	p_ic->line_len = 8 << ibcr.line_len;
 142	p_ic->sz_k = 1 << (ibcr.sz - 1);
 143	p_ic->ver = ibcr.ver;
 144	p_ic->vipt = 1;
 145	p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
 146
 147dc_chk:
 148	p_dc = &cpuinfo_arc700[cpu].dcache;
 149	READ_BCR(ARC_REG_DC_BCR, dbcr);
 150
 151	if (!dbcr.ver)
 152		goto slc_chk;
 153
 154	if (dbcr.ver <= 3) {
 155		BUG_ON(dbcr.config != 2);
 156		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
 157		p_dc->vipt = 1;
 158		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
 159	} else if (dbcr.ver >= 4) {
 160		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
 161		p_dc->vipt = 0;
 162		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
 163	}
 164
 165	p_dc->line_len = 16 << dbcr.line_len;
 166	p_dc->sz_k = 1 << (dbcr.sz - 1);
 167	p_dc->ver = dbcr.ver;
 168
 169slc_chk:
 170	if (is_isa_arcv2())
 171                read_decode_cache_bcr_arcv2(cpu);
 172}
 173
 174/*
 175 * Line Operation on {I,D}-Cache
 176 */
 177
 178#define OP_INV		0x1
 179#define OP_FLUSH	0x2
 180#define OP_FLUSH_N_INV	0x3
 181#define OP_INV_IC	0x4
 182
 183/*
 184 *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
 185 *
 186 * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
 187 * The orig Cache Management Module "CDU" only required paddr to invalidate a
 188 * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
 189 * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
 190 * the exact same line.
 191 *
 192 * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
 193 * paddr alone could not be used to correctly index the cache.
 194 *
 195 * ------------------
 196 * MMU v1/v2 (Fixed Page Size 8k)
 197 * ------------------
 198 * The solution was to provide CDU with these additonal vaddr bits. These
 199 * would be bits [x:13], x would depend on cache-geometry, 13 comes from
 200 * standard page size of 8k.
 201 * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
 202 * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
 203 * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
 204 * represent the offset within cache-line. The adv of using this "clumsy"
 205 * interface for additional info was no new reg was needed in CDU programming
 206 * model.
 207 *
 208 * 17:13 represented the max num of bits passable, actual bits needed were
 209 * fewer, based on the num-of-aliases possible.
 210 * -for 2 alias possibility, only bit 13 needed (32K cache)
 211 * -for 4 alias possibility, bits 14:13 needed (64K cache)
 212 *
 213 * ------------------
 214 * MMU v3
 215 * ------------------
 216 * This ver of MMU supports variable page sizes (1k-16k): although Linux will
 217 * only support 8k (default), 16k and 4k.
 218 * However from hardware perspective, smaller page sizes aggrevate aliasing
 219 * meaning more vaddr bits needed to disambiguate the cache-line-op ;
 220 * the existing scheme of piggybacking won't work for certain configurations.
 221 * Two new registers IC_PTAG and DC_PTAG inttoduced.
 222 * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
 223 */
 224
 225static inline
 226void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
 227			  unsigned long sz, const int op)
 228{
 229	unsigned int aux_cmd;
 230	int num_lines;
 231	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 232
 233	if (op == OP_INV_IC) {
 234		aux_cmd = ARC_REG_IC_IVIL;
 235	} else {
 236		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
 237		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 238	}
 239
 240	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 241	 * and have @paddr - aligned to cache line and integral @num_lines.
 242	 * This however can be avoided for page sized since:
 243	 *  -@paddr will be cache-line aligned already (being page aligned)
 244	 *  -@sz will be integral multiple of line size (being page sized).
 245	 */
 246	if (!full_page) {
 247		sz += paddr & ~CACHE_LINE_MASK;
 248		paddr &= CACHE_LINE_MASK;
 249		vaddr &= CACHE_LINE_MASK;
 250	}
 251
 252	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 253
 254	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
 255	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
 256
 257	while (num_lines-- > 0) {
 258		write_aux_reg(aux_cmd, paddr);
 259		paddr += L1_CACHE_BYTES;
 260	}
 261}
 262
 263/*
 264 * For ARC700 MMUv3 I-cache and D-cache flushes
 265 * Also reused for HS38 aliasing I-cache configuration
 
 
 
 
 266 */
 267static inline
 268void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
 269			  unsigned long sz, const int op)
 270{
 271	unsigned int aux_cmd, aux_tag;
 272	int num_lines;
 273	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 274
 275	if (op == OP_INV_IC) {
 276		aux_cmd = ARC_REG_IC_IVIL;
 277		aux_tag = ARC_REG_IC_PTAG;
 278	} else {
 279		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 280		aux_tag = ARC_REG_DC_PTAG;
 281	}
 282
 283	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 284	 * and have @paddr - aligned to cache line and integral @num_lines.
 285	 * This however can be avoided for page sized since:
 286	 *  -@paddr will be cache-line aligned already (being page aligned)
 287	 *  -@sz will be integral multiple of line size (being page sized).
 288	 */
 289	if (!full_page) {
 290		sz += paddr & ~CACHE_LINE_MASK;
 291		paddr &= CACHE_LINE_MASK;
 292		vaddr &= CACHE_LINE_MASK;
 293	}
 294	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 295
 296	/*
 297	 * MMUv3, cache ops require paddr in PTAG reg
 298	 * if V-P const for loop, PTAG can be written once outside loop
 299	 */
 300	if (full_page)
 301		write_aux_reg(aux_tag, paddr);
 302
 303	/*
 304	 * This is technically for MMU v4, using the MMU v3 programming model
 305	 * Special work for HS38 aliasing I-cache configuratino with PAE40
 306	 *   - upper 8 bits of paddr need to be written into PTAG_HI
 307	 *   - (and needs to be written before the lower 32 bits)
 308	 * Note that PTAG_HI is hoisted outside the line loop
 309	 */
 310	if (is_pae40_enabled() && op == OP_INV_IC)
 311		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
 312
 313	while (num_lines-- > 0) {
 314		if (!full_page) {
 315			write_aux_reg(aux_tag, paddr);
 316			paddr += L1_CACHE_BYTES;
 317		}
 318
 319		write_aux_reg(aux_cmd, vaddr);
 320		vaddr += L1_CACHE_BYTES;
 321	}
 322}
 323
 324/*
 325 * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
 326 * Here's how cache ops are implemented
 327 *
 328 *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
 329 *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
 330 *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
 331 *    respectively, similar to MMU v3 programming model, hence
 332 *    __cache_line_loop_v3() is used)
 333 *
 334 * If PAE40 is enabled, independent of aliasing considerations, the higher bits
 335 * needs to be written into PTAG_HI
 336 */
 337static inline
 338void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
 339			  unsigned long sz, const int cacheop)
 340{
 341	unsigned int aux_cmd;
 342	int num_lines;
 343	const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 344
 345	if (cacheop == OP_INV_IC) {
 346		aux_cmd = ARC_REG_IC_IVIL;
 347	} else {
 348		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
 349		aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 350	}
 351
 352	/* Ensure we properly floor/ceil the non-line aligned/sized requests
 353	 * and have @paddr - aligned to cache line and integral @num_lines.
 354	 * This however can be avoided for page sized since:
 355	 *  -@paddr will be cache-line aligned already (being page aligned)
 356	 *  -@sz will be integral multiple of line size (being page sized).
 357	 */
 358	if (!full_page_op) {
 359		sz += paddr & ~CACHE_LINE_MASK;
 360		paddr &= CACHE_LINE_MASK;
 361	}
 362
 363	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 364
 365	/*
 366	 * For HS38 PAE40 configuration
 367	 *   - upper 8 bits of paddr need to be written into PTAG_HI
 368	 *   - (and needs to be written before the lower 32 bits)
 369	 */
 370	if (is_pae40_enabled()) {
 371		if (cacheop == OP_INV_IC)
 372			/*
 373			 * Non aliasing I-cache in HS38,
 374			 * aliasing I-cache handled in __cache_line_loop_v3()
 375			 */
 376			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
 377		else
 378			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
 379	}
 380
 381	while (num_lines-- > 0) {
 382		write_aux_reg(aux_cmd, paddr);
 383		paddr += L1_CACHE_BYTES;
 384	}
 385}
 386
 387#if (CONFIG_ARC_MMU_VER < 3)
 388#define __cache_line_loop	__cache_line_loop_v2
 389#elif (CONFIG_ARC_MMU_VER == 3)
 390#define __cache_line_loop	__cache_line_loop_v3
 391#elif (CONFIG_ARC_MMU_VER > 3)
 392#define __cache_line_loop	__cache_line_loop_v4
 393#endif
 394
 395#ifdef CONFIG_ARC_HAS_DCACHE
 396
 397/***************************************************************
 398 * Machine specific helpers for Entire D-Cache or Per Line ops
 399 */
 400
 401static inline void __before_dc_op(const int op)
 402{
 403	if (op == OP_FLUSH_N_INV) {
 404		/* Dcache provides 2 cmd: FLUSH or INV
 405		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
 406		 * flush-n-inv is achieved by INV cmd but with IM=1
 407		 * So toggle INV sub-mode depending on op request and default
 408		 */
 409		const unsigned int ctl = ARC_REG_DC_CTRL;
 410		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
 411	}
 412}
 413
 414static inline void __after_dc_op(const int op)
 415{
 416	if (op & OP_FLUSH) {
 417		const unsigned int ctl = ARC_REG_DC_CTRL;
 418		unsigned int reg;
 419
 420		/* flush / flush-n-inv both wait */
 421		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
 422			;
 423
 424		/* Switch back to default Invalidate mode */
 425		if (op == OP_FLUSH_N_INV)
 426			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
 427	}
 428}
 429
 430/*
 431 * Operation on Entire D-Cache
 432 * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
 433 * Note that constant propagation ensures all the checks are gone
 434 * in generated code
 435 */
 436static inline void __dc_entire_op(const int op)
 437{
 438	int aux;
 439
 440	__before_dc_op(op);
 441
 442	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
 443		aux = ARC_REG_DC_IVDC;
 444	else
 445		aux = ARC_REG_DC_FLSH;
 446
 447	write_aux_reg(aux, 0x1);
 448
 449	__after_dc_op(op);
 450}
 451
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 452/* For kernel mappings cache operation: index is same as paddr */
 453#define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
 454
 455/*
 456 * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
 457 */
 458static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
 459				unsigned long sz, const int op)
 460{
 461	unsigned long flags;
 462
 463	local_irq_save(flags);
 464
 465	__before_dc_op(op);
 466
 467	__cache_line_loop(paddr, vaddr, sz, op);
 468
 469	__after_dc_op(op);
 470
 471	local_irq_restore(flags);
 472}
 473
 474#else
 475
 476#define __dc_entire_op(op)
 
 
 477#define __dc_line_op(paddr, vaddr, sz, op)
 478#define __dc_line_op_k(paddr, sz, op)
 479
 480#endif /* CONFIG_ARC_HAS_DCACHE */
 481
 482#ifdef CONFIG_ARC_HAS_ICACHE
 483
 484static inline void __ic_entire_inv(void)
 485{
 486	write_aux_reg(ARC_REG_IC_IVIC, 1);
 487	read_aux_reg(ARC_REG_IC_CTRL);	/* blocks */
 488}
 489
 490static inline void
 491__ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
 492			  unsigned long sz)
 493{
 494	unsigned long flags;
 495
 496	local_irq_save(flags);
 497	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC);
 498	local_irq_restore(flags);
 499}
 500
 501#ifndef CONFIG_SMP
 502
 503#define __ic_line_inv_vaddr(p, v, s)	__ic_line_inv_vaddr_local(p, v, s)
 504
 505#else
 506
 507struct ic_inv_args {
 508	phys_addr_t paddr, vaddr;
 509	int sz;
 510};
 511
 512static void __ic_line_inv_vaddr_helper(void *info)
 513{
 514        struct ic_inv_args *ic_inv = info;
 515
 516        __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
 517}
 518
 519static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
 520				unsigned long sz)
 521{
 522	struct ic_inv_args ic_inv = {
 523		.paddr = paddr,
 524		.vaddr = vaddr,
 525		.sz    = sz
 526	};
 527
 528	on_each_cpu(__ic_line_inv_vaddr_helper, &ic_inv, 1);
 529}
 530
 531#endif	/* CONFIG_SMP */
 532
 533#else	/* !CONFIG_ARC_HAS_ICACHE */
 534
 535#define __ic_entire_inv()
 536#define __ic_line_inv_vaddr(pstart, vstart, sz)
 537
 538#endif /* CONFIG_ARC_HAS_ICACHE */
 539
 540noinline void slc_op(phys_addr_t paddr, unsigned long sz, const int op)
 541{
 542#ifdef CONFIG_ISA_ARCV2
 543	/*
 544	 * SLC is shared between all cores and concurrent aux operations from
 545	 * multiple cores need to be serialized using a spinlock
 546	 * A concurrent operation can be silently ignored and/or the old/new
 547	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
 548	 * below)
 549	 */
 550	static DEFINE_SPINLOCK(lock);
 551	unsigned long flags;
 552	unsigned int ctrl;
 553
 554	spin_lock_irqsave(&lock, flags);
 555
 556	/*
 557	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
 558	 *  - b'000 (default) is Flush,
 559	 *  - b'001 is Invalidate if CTRL.IM == 0
 560	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
 561	 */
 562	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
 563
 564	/* Don't rely on default value of IM bit */
 565	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
 566		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
 567	else
 568		ctrl |= SLC_CTRL_IM;
 569
 570	if (op & OP_INV)
 571		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
 572	else
 573		ctrl &= ~SLC_CTRL_RGN_OP_INV;
 574
 575	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
 576
 577	/*
 578	 * Lower bits are ignored, no need to clip
 579	 * END needs to be setup before START (latter triggers the operation)
 580	 * END can't be same as START, so add (l2_line_sz - 1) to sz
 581	 */
 582	write_aux_reg(ARC_REG_SLC_RGN_END, (paddr + sz + l2_line_sz - 1));
 583	write_aux_reg(ARC_REG_SLC_RGN_START, paddr);
 584
 585	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
 586
 587	spin_unlock_irqrestore(&lock, flags);
 588#endif
 589}
 590
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 591/***********************************************************
 592 * Exported APIs
 593 */
 594
 595/*
 596 * Handle cache congruency of kernel and userspace mappings of page when kernel
 597 * writes-to/reads-from
 598 *
 599 * The idea is to defer flushing of kernel mapping after a WRITE, possible if:
 600 *  -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
 601 *  -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
 602 *  -In SMP, if hardware caches are coherent
 603 *
 604 * There's a corollary case, where kernel READs from a userspace mapped page.
 605 * If the U-mapping is not congruent to to K-mapping, former needs flushing.
 606 */
 607void flush_dcache_page(struct page *page)
 608{
 609	struct address_space *mapping;
 610
 611	if (!cache_is_vipt_aliasing()) {
 612		clear_bit(PG_dc_clean, &page->flags);
 613		return;
 614	}
 615
 616	/* don't handle anon pages here */
 617	mapping = page_mapping(page);
 618	if (!mapping)
 619		return;
 620
 621	/*
 622	 * pagecache page, file not yet mapped to userspace
 623	 * Make a note that K-mapping is dirty
 624	 */
 625	if (!mapping_mapped(mapping)) {
 626		clear_bit(PG_dc_clean, &page->flags);
 627	} else if (page_mapcount(page)) {
 628
 629		/* kernel reading from page with U-mapping */
 630		phys_addr_t paddr = (unsigned long)page_address(page);
 631		unsigned long vaddr = page->index << PAGE_SHIFT;
 632
 633		if (addr_not_cache_congruent(paddr, vaddr))
 634			__flush_dcache_page(paddr, vaddr);
 635	}
 636}
 637EXPORT_SYMBOL(flush_dcache_page);
 638
 639/*
 640 * DMA ops for systems with L1 cache only
 641 * Make memory coherent with L1 cache by flushing/invalidating L1 lines
 642 */
 643static void __dma_cache_wback_inv_l1(phys_addr_t start, unsigned long sz)
 644{
 645	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
 646}
 647
 648static void __dma_cache_inv_l1(phys_addr_t start, unsigned long sz)
 649{
 650	__dc_line_op_k(start, sz, OP_INV);
 651}
 652
 653static void __dma_cache_wback_l1(phys_addr_t start, unsigned long sz)
 654{
 655	__dc_line_op_k(start, sz, OP_FLUSH);
 656}
 657
 658/*
 659 * DMA ops for systems with both L1 and L2 caches, but without IOC
 660 * Both L1 and L2 lines need to be explicitly flushed/invalidated
 661 */
 662static void __dma_cache_wback_inv_slc(phys_addr_t start, unsigned long sz)
 663{
 664	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
 665	slc_op(start, sz, OP_FLUSH_N_INV);
 666}
 667
 668static void __dma_cache_inv_slc(phys_addr_t start, unsigned long sz)
 669{
 670	__dc_line_op_k(start, sz, OP_INV);
 671	slc_op(start, sz, OP_INV);
 672}
 673
 674static void __dma_cache_wback_slc(phys_addr_t start, unsigned long sz)
 675{
 676	__dc_line_op_k(start, sz, OP_FLUSH);
 677	slc_op(start, sz, OP_FLUSH);
 678}
 679
 680/*
 681 * DMA ops for systems with IOC
 682 * IOC hardware snoops all DMA traffic keeping the caches consistent with
 683 * memory - eliding need for any explicit cache maintenance of DMA buffers
 684 */
 685static void __dma_cache_wback_inv_ioc(phys_addr_t start, unsigned long sz) {}
 686static void __dma_cache_inv_ioc(phys_addr_t start, unsigned long sz) {}
 687static void __dma_cache_wback_ioc(phys_addr_t start, unsigned long sz) {}
 688
 689/*
 690 * Exported DMA API
 691 */
 692void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
 693{
 694	__dma_cache_wback_inv(start, sz);
 695}
 696EXPORT_SYMBOL(dma_cache_wback_inv);
 697
 698void dma_cache_inv(phys_addr_t start, unsigned long sz)
 699{
 700	__dma_cache_inv(start, sz);
 701}
 702EXPORT_SYMBOL(dma_cache_inv);
 703
 704void dma_cache_wback(phys_addr_t start, unsigned long sz)
 705{
 706	__dma_cache_wback(start, sz);
 707}
 708EXPORT_SYMBOL(dma_cache_wback);
 709
 710/*
 711 * This is API for making I/D Caches consistent when modifying
 712 * kernel code (loadable modules, kprobes, kgdb...)
 713 * This is called on insmod, with kernel virtual address for CODE of
 714 * the module. ARC cache maintenance ops require PHY address thus we
 715 * need to convert vmalloc addr to PHY addr
 716 */
 717void flush_icache_range(unsigned long kstart, unsigned long kend)
 718{
 719	unsigned int tot_sz;
 720
 721	WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
 722
 723	/* Shortcut for bigger flush ranges.
 724	 * Here we don't care if this was kernel virtual or phy addr
 725	 */
 726	tot_sz = kend - kstart;
 727	if (tot_sz > PAGE_SIZE) {
 728		flush_cache_all();
 729		return;
 730	}
 731
 732	/* Case: Kernel Phy addr (0x8000_0000 onwards) */
 733	if (likely(kstart > PAGE_OFFSET)) {
 734		/*
 735		 * The 2nd arg despite being paddr will be used to index icache
 736		 * This is OK since no alternate virtual mappings will exist
 737		 * given the callers for this case: kprobe/kgdb in built-in
 738		 * kernel code only.
 739		 */
 740		__sync_icache_dcache(kstart, kstart, kend - kstart);
 741		return;
 742	}
 743
 744	/*
 745	 * Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
 746	 * (1) ARC Cache Maintenance ops only take Phy addr, hence special
 747	 *     handling of kernel vaddr.
 748	 *
 749	 * (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
 750	 *     it still needs to handle  a 2 page scenario, where the range
 751	 *     straddles across 2 virtual pages and hence need for loop
 752	 */
 753	while (tot_sz > 0) {
 754		unsigned int off, sz;
 755		unsigned long phy, pfn;
 756
 757		off = kstart % PAGE_SIZE;
 758		pfn = vmalloc_to_pfn((void *)kstart);
 759		phy = (pfn << PAGE_SHIFT) + off;
 760		sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
 761		__sync_icache_dcache(phy, kstart, sz);
 762		kstart += sz;
 763		tot_sz -= sz;
 764	}
 765}
 766EXPORT_SYMBOL(flush_icache_range);
 767
 768/*
 769 * General purpose helper to make I and D cache lines consistent.
 770 * @paddr is phy addr of region
 771 * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc)
 772 *    However in one instance, when called by kprobe (for a breakpt in
 773 *    builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
 774 *    use a paddr to index the cache (despite VIPT). This is fine since since a
 775 *    builtin kernel page will not have any virtual mappings.
 776 *    kprobe on loadable module will be kernel vaddr.
 777 */
 778void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
 779{
 780	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
 781	__ic_line_inv_vaddr(paddr, vaddr, len);
 782}
 783
 784/* wrapper to compile time eliminate alignment checks in flush loop */
 785void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
 786{
 787	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
 788}
 789
 790/*
 791 * wrapper to clearout kernel or userspace mappings of a page
 792 * For kernel mappings @vaddr == @paddr
 793 */
 794void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
 795{
 796	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
 797}
 798
 799noinline void flush_cache_all(void)
 800{
 801	unsigned long flags;
 802
 803	local_irq_save(flags);
 804
 805	__ic_entire_inv();
 806	__dc_entire_op(OP_FLUSH_N_INV);
 807
 808	local_irq_restore(flags);
 809
 810}
 811
 812#ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
 813
 814void flush_cache_mm(struct mm_struct *mm)
 815{
 816	flush_cache_all();
 817}
 818
 819void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
 820		      unsigned long pfn)
 821{
 822	unsigned int paddr = pfn << PAGE_SHIFT;
 823
 824	u_vaddr &= PAGE_MASK;
 825
 826	__flush_dcache_page(paddr, u_vaddr);
 827
 828	if (vma->vm_flags & VM_EXEC)
 829		__inv_icache_page(paddr, u_vaddr);
 830}
 831
 832void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 833		       unsigned long end)
 834{
 835	flush_cache_all();
 836}
 837
 838void flush_anon_page(struct vm_area_struct *vma, struct page *page,
 839		     unsigned long u_vaddr)
 840{
 841	/* TBD: do we really need to clear the kernel mapping */
 842	__flush_dcache_page(page_address(page), u_vaddr);
 843	__flush_dcache_page(page_address(page), page_address(page));
 844
 845}
 846
 847#endif
 848
 849void copy_user_highpage(struct page *to, struct page *from,
 850	unsigned long u_vaddr, struct vm_area_struct *vma)
 851{
 852	void *kfrom = kmap_atomic(from);
 853	void *kto = kmap_atomic(to);
 854	int clean_src_k_mappings = 0;
 855
 856	/*
 857	 * If SRC page was already mapped in userspace AND it's U-mapping is
 858	 * not congruent with K-mapping, sync former to physical page so that
 859	 * K-mapping in memcpy below, sees the right data
 860	 *
 861	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
 862	 * equally valid for SRC page as well
 863	 *
 864	 * For !VIPT cache, all of this gets compiled out as
 865	 * addr_not_cache_congruent() is 0
 866	 */
 867	if (page_mapcount(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
 868		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
 869		clean_src_k_mappings = 1;
 870	}
 871
 872	copy_page(kto, kfrom);
 873
 874	/*
 875	 * Mark DST page K-mapping as dirty for a later finalization by
 876	 * update_mmu_cache(). Although the finalization could have been done
 877	 * here as well (given that both vaddr/paddr are available).
 878	 * But update_mmu_cache() already has code to do that for other
 879	 * non copied user pages (e.g. read faults which wire in pagecache page
 880	 * directly).
 881	 */
 882	clear_bit(PG_dc_clean, &to->flags);
 883
 884	/*
 885	 * if SRC was already usermapped and non-congruent to kernel mapping
 886	 * sync the kernel mapping back to physical page
 887	 */
 888	if (clean_src_k_mappings) {
 889		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
 890		set_bit(PG_dc_clean, &from->flags);
 891	} else {
 892		clear_bit(PG_dc_clean, &from->flags);
 893	}
 894
 895	kunmap_atomic(kto);
 896	kunmap_atomic(kfrom);
 897}
 898
 899void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
 900{
 901	clear_page(to);
 902	clear_bit(PG_dc_clean, &page->flags);
 903}
 904
 905
 906/**********************************************************************
 907 * Explicit Cache flush request from user space via syscall
 908 * Needed for JITs which generate code on the fly
 909 */
 910SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
 911{
 912	/* TBD: optimize this */
 913	flush_cache_all();
 914	return 0;
 915}
 916
 917void arc_cache_init(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 918{
 919	unsigned int __maybe_unused cpu = smp_processor_id();
 920	char str[256];
 921
 922	printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
 923
 924	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
 925		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
 926
 927		if (!ic->ver)
 928			panic("cache support enabled but non-existent cache\n");
 929
 930		if (ic->line_len != L1_CACHE_BYTES)
 931			panic("ICache line [%d] != kernel Config [%d]",
 932			      ic->line_len, L1_CACHE_BYTES);
 933
 934		if (ic->ver != CONFIG_ARC_MMU_VER)
 935			panic("Cache ver [%d] doesn't match MMU ver [%d]\n",
 936			      ic->ver, CONFIG_ARC_MMU_VER);
 937
 938		/*
 939		 * In MMU v4 (HS38x) the alising icache config uses IVIL/PTAG
 940		 * pair to provide vaddr/paddr respectively, just as in MMU v3
 941		 */
 942		if (is_isa_arcv2() && ic->alias)
 943			_cache_line_loop_ic_fn = __cache_line_loop_v3;
 944		else
 945			_cache_line_loop_ic_fn = __cache_line_loop;
 946	}
 947
 948	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
 949		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
 950
 951		if (!dc->ver)
 952			panic("cache support enabled but non-existent cache\n");
 953
 954		if (dc->line_len != L1_CACHE_BYTES)
 955			panic("DCache line [%d] != kernel Config [%d]",
 956			      dc->line_len, L1_CACHE_BYTES);
 957
 958		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
 959		if (is_isa_arcompact()) {
 960			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
 
 961
 962			if (dc->alias && !handled)
 963				panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
 964			else if (!dc->alias && handled)
 
 
 
 965				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
 
 966		}
 967	}
 968
 969	if (is_isa_arcv2() && l2_line_sz && !slc_enable) {
 
 
 970
 971		/* IM set : flush before invalidate */
 972		write_aux_reg(ARC_REG_SLC_CTRL,
 973			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_IM);
 974
 975		write_aux_reg(ARC_REG_SLC_INVALIDATE, 1);
 976
 977		/* Important to wait for flush to complete */
 978		while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
 979		write_aux_reg(ARC_REG_SLC_CTRL,
 980			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_DISABLE);
 981	}
 982
 983	if (is_isa_arcv2() && ioc_exists) {
 984		/* IO coherency base - 0x8z */
 985		write_aux_reg(ARC_REG_IO_COH_AP0_BASE, 0x80000);
 986		/* IO coherency aperture size - 512Mb: 0x8z-0xAz */
 987		write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, 0x11);
 988		/* Enable partial writes */
 989		write_aux_reg(ARC_REG_IO_COH_PARTIAL, 1);
 990		/* Enable IO coherency */
 991		write_aux_reg(ARC_REG_IO_COH_ENABLE, 1);
 992
 
 993		__dma_cache_wback_inv = __dma_cache_wback_inv_ioc;
 994		__dma_cache_inv = __dma_cache_inv_ioc;
 995		__dma_cache_wback = __dma_cache_wback_ioc;
 996	} else if (is_isa_arcv2() && l2_line_sz && slc_enable) {
 997		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
 998		__dma_cache_inv = __dma_cache_inv_slc;
 999		__dma_cache_wback = __dma_cache_wback_slc;
1000	} else {
1001		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
1002		__dma_cache_inv = __dma_cache_inv_l1;
1003		__dma_cache_wback = __dma_cache_wback_l1;
1004	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1005}